Security for builders. AI coding agents run with your full permissions. Hound watches what they actually do.
¶ 01 When you invoke Claude Code or Cursor, it runs as you. Every file you can read, it can read. Every server you can reach, it can reach.
¶ 02 Most of the time, nothing goes wrong. But AI agents now install packages, run scripts, and execute code you've never reviewed. The attack surface is real.
¶ 03 Hound watches every process the agent spawns. The moment something exceeds its scope — a credential read, an unknown connection, a persistence attempt — Hound catches it and shows you exactly what happened.
¶ 04 Blocking rolls out one rule at a time — each proven safe before it can ever stop a process.
If you've ever left an agent running overnight, shipped a PR at 2am, or typed yes to a prompt you half-read — this is for you. Hound is the quiet system in the background that makes that okay.
A Trace is the full story of what an agent did while it had your keys — every file touched, every process spawned, every byte it tried to send.
Not a log dump. A narrative. Readable by the person who wrote the prompt, not just the person who wrote the kernel module.
Share it with a teammate, attach it to a PR, or feed it back to the model. The Trace is the receipt. And it stays yours — local-first, nothing leaves your machine.
No spam. No marketing.
Built for macOS on Apple Silicon. More platforms to come.